R Bootcamp, Module 6: Data manipulation with the tidyverse
January 2019, UC Berkeley
Nima Hejazi (based on materials developed by Kellie Ottoboni, Rochelle Terman, and Chris Krogslund)
Overview
It is often said that 80% of data analysis is spent on the process of cleaning and preparing the data. (Dasu and Johnson, 2003)
Thus before you can even get to doing any sort of sophisticated analysis or plotting, you'll generally first need to:
- Manipulating data frames, e.g., filtering, summarizing, and conducting calculations across groups.
- Tidying data into the appropriate format
There are two competing schools of thought within the R community.
- We should stick to the base R functions to do manipulating and tidying;
tidyverse uses syntax that's unlike base R and is superfluous.
- We should start teaching students to manipulate data using
tidyverse tools because they are straightfoward to use, more readable than base R, and speed up the tidying process.
We'll show you some of the tidyverse tools so you can make an informed decision about whether you want to use base R or these newfangled packages.
Data frame Manipulation using Base R Functions
So far, you've seen the basics of manipulating data frames, e.g. subsetting, merging, and basic calculations. For instance, we can use base R functions to calculate summary statistics across groups of observations, e.g., the mean GDP per capita within each region:
mean(gap[gap$continent == "Africa", "gdpPercap"])
## [1] 2193.755
mean(gap[gap$continent == "Americas", "gdpPercap"])
## [1] 7136.11
mean(gap[gap$continent == "Asia", "gdpPercap"])
## [1] 7902.15
But this isn't ideal because it involves a fair bit of repetition. Repeating yourself will cost you time, both now and later, and potentially introduce some nasty bugs.
Data frame Manipulation using dplyr
Luckily, the dplyr package provides a number of very useful functions for manipulating data frames. These functions will save you time by reducing repetition. As an added bonus, you might even find the dplyr grammar easier to read.
Here we're going to cover 6 of the most commonly used functions as well as using pipes (%>%) to combine them.
select()filter()group_by()summarize()mutate()arrange()
If you have have not installed this package earlier, please do so now:
# NOT run
install.packages('dplyr')
Now let's load the package:
dplyr::select
Imagine that we just received the gapminder dataset, but are only interested in a few variables in it. We could use the select() function to keep only the columns corresponding to variables we select.
year_country_gdp_dplyr <- select(gap, year, country, gdpPercap)
head(year_country_gdp_dplyr)
## year country gdpPercap
## 1 1952 Afghanistan 779.4453
## 2 1957 Afghanistan 820.8530
## 3 1962 Afghanistan 853.1007
## 4 1967 Afghanistan 836.1971
## 5 1972 Afghanistan 739.9811
## 6 1977 Afghanistan 786.1134
If we open up year_country_gdp, we'll see that it only contains the year, country and gdpPercap. This is equivalent to the base R subsetting function:
year_country_gdp_base <- gap[,c("year", "country", "gdpPercap")]
head(year_country_gdp_base)
## year country gdpPercap
## 1 1952 Afghanistan 779.4453
## 2 1957 Afghanistan 820.8530
## 3 1962 Afghanistan 853.1007
## 4 1967 Afghanistan 836.1971
## 5 1972 Afghanistan 739.9811
## 6 1977 Afghanistan 786.1134
We can even check that these two data frames are equivalent:
# checking equivalence: TRUE indicates an exact match between these objects
all.equal(year_country_gdp_dplyr, year_country_gdp_base)
## [1] TRUE
But, as we will see, dplyr makes for much more readable, efficient code because of its pipe operator.
piping with dplyr
Above, we used what's called "normal" grammar, but the strengths of dplyr lie in combining several functions using pipes. Pipes take the input on the left side of the %>% symbol and pass it in as the first argument to the function on the right side. Since the pipe grammar is unlike anything we've seen in R before, let's repeat what we've done above using pipes.
year_country_gdp <- gap %>% select(year, country, gdpPercap)
First we summon the gapminder dataframe and pass it on to the next step using the pipe symbol %>% The second steps is the select() function. In this case we don't specify which data object we use in the call to select() since we've piped it in.
Fun Fact: There is a good chance you have encountered pipes before in the shell. In R, a pipe symbol is %>% while in the shell it is |. But the concept is the same!
dplyr::filter
Now let's say we're only interested in African countries. We can combine select and filter to select only the observations where continent is Africa.
year_country_gdp_africa <- gap %>%
filter(continent == "Africa") %>%
select(year,country,gdpPercap)
As with last time, first we pass the gapminder data frame to the filter() function, then we pass the filtered version of the gapminder data frame to the select() function.
To clarify, both the select and filter functions subsets the data frame. The difference is that select extracts certain columns, while filter extracts certain rows.
Note: The order of operations is very important in this case. If we used 'select' first, filter would not be able to find the variable continent since we would have removed it in the previous step.
dplyr Calculations Across Groups
A common task you'll encounter when working with data is running calculations on different groups within the data. For instance, what if we wanted to calculate the mean GDP per capita for each continent?
In base R, you would have to run the mean() function for each subset of data:
mean(gap$gdpPercap[gap$continent == "Africa"])
## [1] 2193.755
mean(gap$gdpPercap[gap$continent == "Americas"])
## [1] 7136.11
mean(gap$gdpPercap[gap$continent == "Asia"])
## [1] 7902.15
mean(gap$gdpPercap[gap$continent == "Europe"])
## [1] 14469.48
mean(gap$gdpPercap[gap$continent == "Oceania"])
## [1] 18621.61
That's a lot of repetition! To make matters worse, what if we wanted to add these values to our original data frame as a new column? We would have to write something like this:
gap$mean.continent.GDP <- NA
gap$mean.continent.GDP[gap$continent == "Africa"] <- mean(gap$gdpPercap[gap$continent == "Africa"])
gap$mean.continent.GDP[gap$continent == "Americas"] <- mean(gap$gdpPercap[gap$continent == "Americas"])
gap$mean.continent.GDP[gap$continent == "Asia"] <- mean(gap$gdpPercap[gap$continent == "Asia"])
gap$mean.continent.GDP[gap$continent == "Europe"] <- mean(gap$gdpPercap[gap$continent == "Europe"])
gap$mean.continent.GDP[gap$continent == "Oceania"] <- mean(gap$gdpPercap[gap$continent == "Oceania"])
You can see how this can get pretty tedious, especially if we want to calculate more complicated or refined statistics. We could use loops or apply functions, but these can be difficult, slow, or error-prone.
dplyr split-apply-combine
The abstract problem we're encountering here is know as "split-apply-combine":
We want to split our data into groups (in this case continents), apply some calculations on each group, then combine the results together afterwards.
Module 4 gave some ways to do split-apply-combine type stuff using the apply family of functions, but those are error prone and messy.
Luckily, dplyr offers a much cleaner, straight-forward solution to this problem.
# remove this column -- there are two easy ways!
gap <- gap %>% select(-mean.continent.GDP)
# OR
gap$mean.continent.GDP <- NULL
dplyr::group_by
We've already seen how filter() can help us select observations that meet certain criteria (in the above: continent == "Europe"). More helpful, however, is the group_by() function, which will essentially use every unique criteria that we could have used in filter().
A grouped_df can be thought of as a list where each item in the list is a data.frame which contains only the rows that correspond to the a particular value continent (at least in the example above).
dplyr::summarize
group_by() on its own is not particularly interesting. It's much more exciting used in conjunction with the summarize() function. This will allow use to create new variable(s) by applying transformations to variables in each of the continent-specific data frames. In other words, using the group_by() function, we split our original data frame into multiple pieces, which we then apply summary functions to (e.g. mean() or sd()) within summarize(). The output is a new data frame reduced in size, with one row per group.
gdp_bycontinents <- gap %>%
group_by(continent) %>%
summarize(mean_gdpPercap = mean(gdpPercap))
head(gdp_bycontinents)
## # A tibble: 5 x 2
## continent mean_gdpPercap
## <chr> <dbl>
## 1 Africa 2194.
## 2 Americas 7136.
## 3 Asia 7902.
## 4 Europe 14469.
## 5 Oceania 18622.
That allowed us to calculate the mean gdpPercap for each continent. But it gets even better -- the function group_by() allows us to group by multiple variables. Let's group by year and continent.
gdp_bycontinents_byyear <- gap %>%
group_by(continent, year) %>%
summarize(mean_gdpPercap = mean(gdpPercap))
head(gdp_bycontinents_byyear)
## # A tibble: 6 x 3
## # Groups: continent [1]
## continent year mean_gdpPercap
## <chr> <int> <dbl>
## 1 Africa 1952 1253.
## 2 Africa 1957 1385.
## 3 Africa 1962 1598.
## 4 Africa 1967 2050.
## 5 Africa 1972 2340.
## 6 Africa 1977 2586.
That is already quite powerful, but it gets even better! You're not limited to defining 1 new variable in summarize().
gdp_pop_bycontinents_byyear <- gap %>%
group_by(continent, year) %>%
summarize(mean_gdpPercap = mean(gdpPercap),
sd_gdpPercap = sd(gdpPercap),
mean_pop = mean(pop),
sd_pop = sd(pop))
head(gdp_pop_bycontinents_byyear)
## # A tibble: 6 x 6
## # Groups: continent [1]
## continent year mean_gdpPercap sd_gdpPercap mean_pop sd_pop
## <chr> <int> <dbl> <dbl> <dbl> <dbl>
## 1 Africa 1952 1253. 983. 4570010. 6317450.
## 2 Africa 1957 1385. 1135. 5093033. 7076042.
## 3 Africa 1962 1598. 1462. 5702247. 7957545.
## 4 Africa 1967 2050. 2848. 6447875. 8985505.
## 5 Africa 1972 2340. 3287. 7305376. 10130833.
## 6 Africa 1977 2586. 4142. 8328097. 11585184.
dplyr::mutate
What if we wanted to add these values to our original data frame instead of creating a new object? For this, we can use the mutate() function, which is similar to summarize() except it creates new variables to the same data frame that you pass into it.
gap_with_extra_vars <- gap %>%
group_by(continent, year) %>%
mutate(mean_gdpPercap = mean(gdpPercap),
sd_gdpPercap = sd(gdpPercap),
mean_pop = mean(pop),
sd_pop = sd(pop))
head(gap_with_extra_vars)
## # A tibble: 6 x 10
## # Groups: continent, year [6]
## country year pop continent lifeExp gdpPercap mean_gdpPercap
## <chr> <int> <dbl> <chr> <dbl> <dbl> <dbl>
## 1 Afghanistan 1952 8425333 Asia 28.8 779. 5195.
## 2 Afghanistan 1957 9240934 Asia 30.3 821. 5788.
## 3 Afghanistan 1962 10267083 Asia 32.0 853. 5729.
## 4 Afghanistan 1967 11537966 Asia 34.0 836. 5971.
## 5 Afghanistan 1972 13079460 Asia 36.1 740. 8187.
## 6 Afghanistan 1977 14880372 Asia 38.4 786. 7791.
## # ... with 3 more variables: sd_gdpPercap <dbl>, mean_pop <dbl>,
## # sd_pop <dbl>
We can use also use mutate() to create new variables prior to (or even after) summarizing information.
gdp_pop_bycontinents_byyear <- gap %>%
mutate(gdp_billion = gdpPercap*pop/10^9) %>%
group_by(continent, year) %>%
summarize(mean_gdpPercap = mean(gdpPercap),
sd_gdpPercap = sd(gdpPercap),
mean_pop = mean(pop),
sd_pop = sd(pop),
mean_gdp_billion = mean(gdp_billion),
sd_gdp_billion = sd(gdp_billion))
head(gdp_pop_bycontinents_byyear)
## # A tibble: 6 x 8
## # Groups: continent [1]
## continent year mean_gdpPercap sd_gdpPercap mean_pop sd_pop
## <chr> <int> <dbl> <dbl> <dbl> <dbl>
## 1 Africa 1952 1253. 983. 4570010. 6317450.
## 2 Africa 1957 1385. 1135. 5093033. 7076042.
## 3 Africa 1962 1598. 1462. 5702247. 7957545.
## 4 Africa 1967 2050. 2848. 6447875. 8985505.
## 5 Africa 1972 2340. 3287. 7305376. 10130833.
## 6 Africa 1977 2586. 4142. 8328097. 11585184.
## # ... with 2 more variables: mean_gdp_billion <dbl>, sd_gdp_billion <dbl>
dplyr::arrange
As a last step, let's say we want to sort the rows in our data frame according to values in a certain column. We can use the arrange() function to do this. For instance, let's organize our rows by year (recent first), and then by continent.
gap_with_extra_vars <- gap %>%
group_by(continent, year) %>%
mutate(mean_gdpPercap = mean(gdpPercap),
sd_gdpPercap = sd(gdpPercap),
mean_pop = mean(pop),
sd_pop = sd(pop)) %>%
arrange(desc(year), continent)
head(gap_with_extra_vars)
## # A tibble: 6 x 10
## # Groups: continent, year [1]
## country year pop continent lifeExp gdpPercap mean_gdpPercap
## <chr> <int> <dbl> <chr> <dbl> <dbl> <dbl>
## 1 Algeria 2007 33333216 Africa 72.3 6223. 3089.
## 2 Angola 2007 12420476 Africa 42.7 4797. 3089.
## 3 Benin 2007 8078314 Africa 56.7 1441. 3089.
## 4 Botswana 2007 1639131 Africa 50.7 12570. 3089.
## 5 Burkina Faso 2007 14326203 Africa 52.3 1217. 3089.
## 6 Burundi 2007 8390505 Africa 49.6 430. 3089.
## # ... with 3 more variables: sd_gdpPercap <dbl>, mean_pop <dbl>,
## # sd_pop <dbl>
dplyr Take-aways
- Human readable: the function names describe the action being done
- Piping: chain functions in a step-by-step way, rather than nesting
# without pipes:
gap_with_extra_vars <- arrange(
mutate(
group_by(gap, continent, year),
mean_gdpPercap = mean(gdpPercap)
),
desc(year), continent)
- Facilitates split-apply-combine manipulations
Tidying Data
Even before we conduct analysis or calculations, we need to put our data into the correct format. The goal here is to rearrange a messy dataset into one that is tidy
The two most important properties of tidy data are:
- Each column is a variable.
- Each row is an observation.
Tidy data is easier to work with, because you have a consistent way of referring to variables (as column names) and observations (as row indices). It then becomes easy to manipulate, visualize, and model.
For more on the concept of tidy data, read Hadley Wickham's paper here
Tidying Data/Wide vs. Long Formats
"Tidy datasets are all alike but every messy dataset is messy in its own way." – Hadley Wickham
Tabular datasets can be arranged in many ways. For instance, consider the data below. Each data set displays information on heart rate observed in individuals across 3 different time periods. But the data are organized differently in each table.
wide <- data.frame(
name = c("Wilbur", "Petunia", "Gregory"),
time1 = c(67, 80, 64),
time2 = c(56, 90, 50),
time3 = c(70, 67, 101)
)
wide
## name time1 time2 time3
## 1 Wilbur 67 56 70
## 2 Petunia 80 90 67
## 3 Gregory 64 50 101
long <- data.frame(
name = c("Wilbur", "Petunia", "Gregory", "Wilbur", "Petunia", "Gregory", "Wilbur", "Petunia", "Gregory"),
time = c(1, 1, 1, 2, 2, 2, 3, 3, 3),
heartrate = c(67, 80, 64, 56, 90, 50, 70, 67, 10)
)
long
## name time heartrate
## 1 Wilbur 1 67
## 2 Petunia 1 80
## 3 Gregory 1 64
## 4 Wilbur 2 56
## 5 Petunia 2 90
## 6 Gregory 2 50
## 7 Wilbur 3 70
## 8 Petunia 3 67
## 9 Gregory 3 10
Question: Which one of these do you think is the tidy format?
Answer: The first dataframe (the "wide" one) would not be considered tidy because values (i.e., heartrate) are spread across multiple columns.
We often refer to these different structurs as "long" vs. "wide" formats. In the "long" format, you usually have 1 column for the observed variable and the other columns are ID variables.
For the "wide" format each row is often a site/subject/patient and you have multiple observation variables containing the same type of data. These can be either repeated observations over time, or observation of multiple variables (or a mix of both). In the above case, we had the same kind of data (heart rate) entered across 3 different columns, corresponding to three different time periods.
You may find data input may be simpler and some programs/functions may prefer the "wide" format. However, many of R’s functions have been designed assuming you have "long" format data.
Tidying the Gapminder Data
Lets look at the structure of our original gapminder dataframe:
## country year pop continent lifeExp gdpPercap
## 1 Afghanistan 1952 8425333 Asia 28.801 779.4453
## 2 Afghanistan 1957 9240934 Asia 30.332 820.8530
## 3 Afghanistan 1962 10267083 Asia 31.997 853.1007
## 4 Afghanistan 1967 11537966 Asia 34.020 836.1971
## 5 Afghanistan 1972 13079460 Asia 36.088 739.9811
## 6 Afghanistan 1977 14880372 Asia 38.438 786.1134
Question: Is this data frame wide or long?
Answer: This data frame is somewhere in between the purely 'long' and 'wide' formats. We have 3 "ID variables" (continent, country, year) and 3 "Observation variables" (pop, lifeExp, gdpPercap).
Despite not having ALL observations in 1 column, this intermediate format makes sense given that all 3 observation variables have different units. As we have seen, many of the functions in R are often vector based, and you usually do not want to do mathematical operations on values with different units.
On the other hand, there are some instances in which a purely long or wide format is ideal (e.g. plotting). Likewise, sometimes you'll get data on your desk that is poorly organized, and you'll need to reshape it.
tidyr
Thankfully, the tidyr package will help you efficiently transform your data regardless of original format.
# Install the "tidyr" package (only necessary one time)
# install.packages("tidyr") # Not Run
# Load the "tidyr" package (necessary every new R session)
library(tidyr)
tidyr::gather
Until now, we've been using the nicely formatted original gapminder data set. This data set is not quite wide and not quite long -- it's something in the middle, but "real" data (i.e., our own research data) will never be so well organized. Here let's start with the wide format version of the gapminder data set.
gap_wide <- read.csv("../data/gap_wide.csv", stringsAsFactors = FALSE)
## Warning in file(file, "rt"): cannot open file '../data/gap_wide.csv': No
## such file or directory
## Error in file(file, "rt"): cannot open the connection
## Error in head(gap_wide): object 'gap_wide' not found
The first step towards getting our nice intermediate data format is to first convert from the wide to the long format. The function gather() will 'gather' the observation variables into a single variable. This is sometimes called "melting" your data, because it melts the table from wide to long. Those data will be melted into two variables: one for the variable names, and the other for the variable values.
gap_long <- gap_wide %>%
gather(obstype_year, obs_values, 3:38)
## Error in eval(lhs, parent, parent): object 'gap_wide' not found
## Error in head(gap_long): object 'gap_long' not found
Notice that we put 3 arguments into the gather() function:
- the name the new column for the new ID variable (
obstype_year),
- the name for the new amalgamated observation variable (
obs_value),
- the indices of the old variables (
3:38, signalling columns 3 through 38) that we want to gather into one variable. Notice that we don't want to melt down columns 1 and 2, as these are considered "ID" variables.
tidyr::select
If there are a lot of columns or they're named in a consistent pattern, we might not want to select them using the column numbers. It'd be easier to use some information contained in the names themselves. We can select variables using:
- variable indices
- variable names (without quotes)
x:z to select all variables between x and z-y to exclude ystarts_with(x, ignore.case = TRUE): all names that starts with xends_with(x, ignore.case = TRUE): all names that ends with xcontains(x, ignore.case = TRUE): all names that contain x
See the select() function in dplyr for more options.
For instance, here we do the same gather operation with (1) the starts_with function, and (2) the - operator:
# with the starts_with() function
gap_long <- gap_wide %>%
gather(obstype_year, obs_values, starts_with('pop'),
starts_with('lifeExp'), starts_with('gdpPercap'))
## Error in eval(lhs, parent, parent): object 'gap_wide' not found
## Error in head(gap_long): object 'gap_long' not found
# with the - operator
gap_long <- gap_wide %>%
gather(obstype_year, obs_values, -continent, -country)
## Error in eval(lhs, parent, parent): object 'gap_wide' not found
## Error in head(gap_long): object 'gap_long' not found
However you choose to do it, notice that the output collapses all of the measure variables into two columns: one containing new ID variable, the other containing the observation value for that row.
tidyr::separate
You'll notice that in our long dataset, obstype_year actually contains 2 pieces of information, the observation type (pop, lifeExp, or gdpPercap) and the year.
We can use the separate() function to split the character strings into multiple variables:
gap_long_sep <- gap_long %>%
separate(obstype_year, into = c('obs_type','year'), sep = "_") %>%
mutate(year = as.integer(year))
## Error in eval(lhs, parent, parent): object 'gap_long' not found
## Error in head(gap_long_sep): object 'gap_long_sep' not found
If you didn't use tidyr to do this, you'd have to use the strsplit function and use multiple lines of code to replace the column in gap_long with two new columns. This solution is much cleaner.
tidyr::spread
The opposite of gather() is spread(). It spreads our observation variables back out to make a wider table. We can use this function to spread our gap_long() to the original "medium" format.
gap_medium <- gap_long_sep %>%
spread(obs_type, obs_values)
## Error in eval(lhs, parent, parent): object 'gap_long_sep' not found
## Error in head(gap_medium): object 'gap_medium' not found
All we need is some quick fixes to make this dataset identical to the original gap dataset:
gap <- read.csv("../data/gapminder-FiveYearData.csv")
head(gap_medium)
## Error in head(gap_medium): object 'gap_medium' not found
## country year pop continent lifeExp gdpPercap
## 1 Afghanistan 1952 8425333 Asia 28.801 779.4453
## 2 Afghanistan 1957 9240934 Asia 30.332 820.8530
## 3 Afghanistan 1962 10267083 Asia 31.997 853.1007
## 4 Afghanistan 1967 11537966 Asia 34.020 836.1971
## 5 Afghanistan 1972 13079460 Asia 36.088 739.9811
## 6 Afghanistan 1977 14880372 Asia 38.438 786.1134
# rearrange columns
gap_medium <- gap_medium[,names(gap)]
## Error in eval(expr, envir, enclos): object 'gap_medium' not found
## Error in head(gap_medium): object 'gap_medium' not found
# arrange by country, continent, and year
gap_medium <- gap_medium %>%
arrange(country,continent,year)
## Error in eval(lhs, parent, parent): object 'gap_medium' not found
## Error in head(gap_medium): object 'gap_medium' not found
Breakout
dplyr
Use dplyr to create a data frame containing the median lifeExp for each continent
Use dplyr to add a column to the gapminder dataset that contains the total population of the continent of each observation in a given year. For example, if the first observation is Afghanistan in 1952, the new column would contain the population of Asia in 1952.
Use dplyr to add a column called gdpPercap_diff that contains the difference between the observation's gdpPercap and the mean gdpPercap of the continent in that year. Arrange the dataframe by the column you just created, in descending order (so that the relatively richest country/years are listed first)
tidyr
- Subset the results from question #3 to select only the
country, year, and gdpPercap_diff columns. Use tidyr put it in wide format so that countries are rows and years are columns.
Hint: you'll probably see a message about a missing grouping variable. If you don't want continent included, you can pass the output of problem 3 through ungroup() to get rid of the continent information.
